Modelpublic

cescox/Qwen3.6-35B-A3B-Abliterated-Heretic-Hebrus-GGUF

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

1likes67downloads

Model Card

Qwen3.6-35B-A3B-Abliterated-Heretic — Hebrus GGUF

This repository holds a Hebrus-native GGUF build of the community Heretic-abliterated Qwen3.6-35B-A3B, dedicated to `Hebrus`, the inference engine created by **andreaborio** as a fork of Salvatore Sanfilippo's `antirez/ds4`. Its routed MoE weights are embedded in the GGUF as a self-describing ds4.expert_major.v2 store, the same format used by `andreaborio/Qwen3.6-35B-A3B-Hebrus-GGUF`, the non-abliterated reference release this file is a drop-in replacement for (same chat template and tokenizer config, different — abliterated — weights).

Uncensored model: this build's refusal behavior has been removed at the weight level using the community Heretic technique. It is materially more willing to comply with a wide range of requests, including harmful ones, than the base Qwen3.6-35B-A3B. Use accordingly.

This is a Hebrus-native GGUF extension, not a portable GGUF layout. Generic GGUF loaders (stock llama.cpp, Ollama, LM Studio, etc.) cannot reconstruct or execute its routed weights — the expert tensors are stored as MLX-compatible affine 4-bit groups, which the standard GGUF spec does not define. If you need a portable GGUF of this fine-tune, use the source repos in Provenance below instead.

Provenance


Base model	`Qwen/Qwen3.6-35B-A3B` — Apache 2.0, Qwen (Alibaba), MoE, 35B total / 3B active params, 40 layers, 256 experts (8 routed + 1 shared)
Abliteration	`Youssofal/Qwen3.6-35B-A3B-Abliterated-Heretic-BF16` — "Heretic" technique (refusal-direction removal from `embed_tokens`, `o_proj`, `down_proj`)
Community repack (source of this build)	`intelservice77/Qwen3.6-35B-A3B-Abliterated-Heretic-BF16-Q4_K_S-GGUF`
Runtime / format	Hebrus, `ds4.expert_major.v2` (ExpertMajor v2 / MLX-affine4-g64)

intelservice77's Q4KS repack used a standard llama-quantize recipe that doesn't respect Hebrus's per-tensor-category type requirements (exact f32 for norms/small SSM params, exact q4k only for routed-expert weights, f16-or-q80 elsewhere). This build fixes that, then converts the corrected canonical GGUF into ExpertMajor v2 / MLX-affine4-g64:

1.Tensor-type fixes: swapped in bytes from the community's Q80 release (same abliterated weights, higher precision — no extra quantization loss) for everything that needed f16-or-q80, borrowed SSM gating params from the unmodified base model (abliteration doesn't touch those), and re-quantized the handful of expert tensors that still needed it.
2.ExpertMajor v2 repack: converted against a bit-depth-normalized MLX 4-bit source of the same fine-tune, using Hebrus's ds4-expert-major.py.

Verified with ds4 --inspect: tensor types and counts identical to the official Hebrus base build. Abliteration behavior confirmed directly: lockpicking/pentesting prompts get direct answers; extreme-content prompts (e.g. explosives) still show partial residual refusal, consistent with Heretic not uniformly clearing every refusal category.

Artifact

Property	Value
File	`Qwen3.6-35B-A3B-Abliterated-Heretic-DS4-ExpertMajor-v2-MLX-Affine4-G64.gguf`
Size	20,808,566,720 bytes (19.38 GiB)
SHA-256	`5a913e1a6cac283fe6cc0fd93b0e48d0a1cd38321810dc822e7a3bad354c037b`
Container	GGUF v3 with one embedded `ds4.expert_major.v2` store
Routed geometry	40 routed layers, 256 experts per layer
Routed storage	MLX-compatible affine 4-bit groups: 32 packed bytes + BF16 scale + BF16 bias per 64 weights
Qualified hardware	Apple Silicon Metal, at least 16 GiB unified memory

Run with Hebrus

Use `andreaborio/hebrus` main at commit `d61a6d73f5c38e92e433beb9e404d06d79b153b1` or a compatible descendant — verified against exactly this build. Earlier checkouts (including the 73a332f cited as the ExpertMajor v2 baseline) hit a real bug, "Qwen Metal batched prefill failed at position 0", reproduced identically against the official non-abliterated release and fixed in this commit.

bash

git clone https://github.com/andreaborio/hebrus.git
cd hebrus
make -j

./hebrus \
  -m /absolute/path/to/Qwen3.6-35B-A3B-Abliterated-Heretic-DS4-ExpertMajor-v2-MLX-Affine4-G64.gguf \
  --ctx 8192

The embedded ExpertMajor v2 store, Metal backend, and residency policy are selected automatically. AUTO uses SSD streaming when the resident budget doesn't fit; the admission floor is 16 GiB unified memory.

Measured performance

Measured on an Apple M1 Pro, 32 GiB, SSD-streamed experts:

Metric	Result
Prefill (5,664-token prompt)	~238 t/s
Decode (after SSD expert-cache warmup)	~15–18 t/s
Long-context retrieval (~5.7k tokens, inserted passphrase)	Correct, no degradation from re-quantization
Multi-step reasoning	Correct, self-verified

These are bounded local measurements; context, output length, thermals, storage, and memory pressure affect results. See `andreaborio/Qwen3.6-35B-A3B-Hebrus-GGUF` for measurements on Apple M5 Pro hardware with the non-abliterated release.

License and attribution

Runtime. andreaborio created and maintains Hebrus, the inference engine this GGUF is built for, including the ExpertMajor v2 / MLX-affine4-g64 format and tooling used throughout this build.

Model provenance.

—Qwen team (Alibaba) — base model, released under Apache 2.0. See the upstream model card for architecture, training, safety, license, and attribution details.
—Youssofal — Heretic abliteration of Qwen3.6-35B-A3B.
—intelservice77 — original Q4KS GGUF repack.